Prognosis and health management play an important role in the control of costs associated with the operation of large industrial equipment. By properly comprehending hardware degradation and accurately predicting the remaining useful life of such equipment, we can significantly lower operational costs by reducing asset downtime and maintenance interventions. However, complex interactions between operational conditions and component capability make accurately modeling damage accumulation for large fleets a daunting task. Unforeseen factors such as aggressive missions introduced by operators, exposure to harsh environments, manufacturing issues, amongst many others, can lead to large discrepancies between predicted and observed useful life. Motivated by the growing availability of data and computational power as well as the advances in hybrid modeling frameworks, capable of merging elements of physics, machine learning, and statistical learning, in this dissertation, we focus on the development of novel approaches to minimize the impact of unforeseen factors in fleet management. In this dissertation, we focus on the challenges of accounting for the impacts of such unforeseen factors on two specific stages of a component service life; early-file and end-life. Two numerical case studies are derived to emulate two common issues in fleet life management; manufacturing issues leading to an infant mortality problem, and unexpected exposure to harsher environments by operators, accelerating wear-out and significantly reducing component's useful life. In the first analysis, two key aspects in a prognosis and health management perspective are addressed; detecting the emerging issue (i.e., the infant mortality problem), and the evaluation of risk mitigation procedures to minimize/mitigate its effects on the overall fleet reliability. Bayesian networks implementing physics-based models are used to model the fleet unreliability and assist in the quantification of the infant mortality impact on the fleet useful life. Additionally, steps to adapted the derived Bayesian networks to assist in the evaluation of possible mitigation approaches to minimize the impacts of fleet-wide early life problems are presented. Concerning the wear-out analysis, a civil aviation case study is derived, in which an aircraft fleet mainly operates in coastal routes, significantly increasing its exposure to saline corrosion. These conditions lead to accelerated degradation of the aircraft wing panels due to the combined effects of corrosion and mechanical fatigue. Such corrosive conditions are not accounted for by the fleet prognosis model generating a significant epistemic uncertainty (i.e., a missing physics issue). To address this issue, we proposed hybrid recurrent neural network modules to compensate for the model-form uncertainty. In the formulated neural network cell, well-understood aspects of the degradation mechanism are addressed by a physics-based model, while data-driven models are trained to account for the missing physics effects. After proper training, the hybrid neural network can compensate for the unaccounted effects in the model damage forecast and generates accurate predictions to assist in the fleet prognosis analysis. Obtained results illustrate the capabilities of the proposed frameworks in compensating for the considered unforeseen factors impacts in fleet management. Additionally, the obtained results have prominently shown the significance and importance of properly account for such factors on fleet prognosis and how these factors can drastically hinder engineers' ability to properly perform prognosis and health management analysis.
If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu
Doctor of Philosophy (Ph.D.)
College of Engineering and Computer Science
Mechanical and Aerospace Engineering
Length of Campus-only Access
Doctoral Dissertation (Open Access)
De Piemonte Dourado, Arinan, "Model-Form Uncertainty Quantification in Prognosis and Fleet Management with Physics-Informed Neural Networks" (2021). Electronic Theses and Dissertations, 2020-. 491.